Top researchers from Anthropic and Google DeepMind have left for an AI startup called Core Automation. Illustration by Avishek Das/SOPA Images/LightRocket via Getty Images A new AI startup, Core Automation, has entered the chat. Started by an ex-OpenAI researcher, it has "nerdsniped" top talent from Anthropic and Google DeepMind. From day one, the startup has branded itself as "the world's most automated AI lab." A new AI startup called Core Automation, founded by an ex-OpenAI researcher, is snatching top talent from Anthropic and Google DeepMind. On Tuesday, Core Automation wrote in its first X post that it is "building the world's most automated AI lab." "Our objective: systems that optimize and automate work, starting with research itself," the company wrote in the post. Jerry Tworek, a former vice president at OpenAI, lists himself as the CEO and cofounder of Core Automation on his X bio. In an X post on Tuesday, Anthropic researcher Rohan Anil said he left the company after Tworek "nerdsniped" him. "Ok I did leave anthropic, a few weeks ago, it was one of the best places to work for a researcher," Anil, who also worked in Google DeepMind, wrote. "Jerry Tworek nerdsniped me into starting this with him and others." Anmol Gulati, a research scientist from Google DeepMind working on Gemini, said in a post that he was "starting something new with some exceptional people." "I've increasingly felt that the current research paradigm — scaling models, data, and static deployment won't get us all the way," Gulati wrote. "We believe the next phase comes from something different: new learning algorithms, architectures beyond today's stack, and systems that automate the process of building itself." Joanne Jang, whose LinkedIn says that she was a general manager at OpenAI from December 2021 to April 2026, has written in her X bio, "trying to automate myself @coreautoai." Others who have shared that they've joined the team include former Google DeepMind staffers Ehsan Amid and Avery Lamp, OpenAI's ex-head of people, Julia Villagra, and Sai Surya Duvvuri, a former research intern at Google and Meta. On its website, Core Automation wrote that its team consists of people who have "helped build frontier models" and "influential architectures." Big Tech to AI startups It's not the first time top AI researchers have left big labs for startups. Yann LeCun, formerly Meta's chief AI scientist, left the company to start his AI startup Advanced Machine Intelligence Labs, also known as AMI Labs. The startup focuses on developing world models — AI systems that better understand and reflect the real world. AMI Labs' approach departs from Meta's focus on commercially driven model development and scaling. Last year, tech giants were battling over top AI talent, offering multibillion-dollar acquihires and massive pay packages. Startups were also active players in the talent war, offering competitive salaries and equity packages, as well as the unique impact and ownership that come with working at a smaller company. Shawn Thorne, managing director at executive search firm True Search, told Business Insider last year that base salaries at startups rose rapidly as they compete to attract AI talent. Equity is "the big factor" helping offset the "opportunity cost" for top researchers or engineers who might otherwise choose to start their own ventures, he said. To sweeten the deal, startups also offer additional incentives such as cofounder titles, access to compute, and time for independent research, Thorne added. Read the original article on Business Insider

Kelly Benthall and her husband retired and have been traveling around the world. Provided by Kelly Benthall After retiring at 53, Kelly Benthall and her husband began traveling around the world. What began as a search for freedom became an unexpected lesson in decision fatigue. Now, in year two, they've found a new way to structure their travel. When I retired early, I assumed the hard part was over. No more meetings. No more deadlines imposed by someone else. My husband and I began traveling in long arcs, staying at least a month in each place. We would unpack, settle in, and let daily life take over. Some weeks were quiet and local; other times we took short trips before returning to reset. When we left our jobs in August 2024, our goal was freedom. What I didn't anticipate was how tiring that would become — not because anything was wrong, but because freedom came with more decisions than I was used to making. After about a year and a half of traveling across 14 countries, most of what I learned came in the last six months. Every day required dozens of small choices: routes, weather, how far was too far, when a short trip became too much. Even rest became something I had to justify. Benthall on a hike near her monthlong home base in Christchurch, New Zealand. Provided by Kelly Benthall Nothing was actually wrong My husband and I were healthy, financially secure, and in beautiful places. This was the life I'd imagined. But by the end of the first year, something felt off. What wore me down wasn't any single experience. It was the accumulation of constant, low-level decisions. Looking back, part of the problem was overconfidence. Early on, we planned the big things — a cave hotel in Matera, a fast-paced route across England, dive recertification in Mauritius. Later, in places like New Zealand, we assumed we could wing it. That worked — until it didn't In Lake Tekapo, we missed the hot pools and stargazing because we hadn't booked ahead. We drove around in the dark instead — technically there, but missing the experience we came for. Moments like that pushed us to move faster, stacking "must-dos" closer together. Soon, the days started competing with each other. I felt it most during what should have been a highlight: landing by helicopter on New Zealand's Earnslaw Burn Glacier. It was vast and cinematic, with waterfalls spilling from rock faces and ice carved over centuries. I'd been looking forward to it for weeks. But standing there, I felt almost nothing. It wasn't the helicopter. It was what came before and after: jet boating, Milford Sound, the TranzAlpine train. Each was unforgettable on its own, but together they blurred. After the helicopter landed on a snowfield near Earnslaw Burn in New Zealand's South Island, she realized she was stacking "must-dos." Provided by Kelly Benthall When everything competes for attention — logistics, decisions, once-in-a-lifetime moments — something gets diluted. I started noticing how often I was deciding. Freedom had quietly turned into constant optimization. My definition of luxury began to change For most of my career, I worked in career in oil and gas, making decisions all day. Most of them small, many of them reversible. The structure of work — meetings, timelines, expectations — meant I wasn't deciding everything from scratch. At work, when a decision felt urgent, I learned to wait. I'd step out of the room, delay the call, let the day pass. Almost always, something shifted, and the decision became clearer. Without those guardrails, the decisions didn't disappear. They multiplied. Provided by Kelly Benthall What I needed wasn't more freedom, but some structure. I didn't need to control everything, but I did want to reduce how much I had to think about. The difference was immediate when we slowed down. We stayed longer, returned to the same café, let the weather set the pace, and said no to side trips that required too much planning. The irony is that I'd chased early retirement to gain control over my time. What I learned instead is that control doesn't come from endless options. It comes from designing your life to require less from you. The most luxurious thing travel gave me wasn't adventure, it was fewer decisions. Read the original article on Business Insider

AI was used to create this 'Elder' as a provocation to research participants. Relational Futures Much of the current conversation about AI assumes uptake is inevitable, more technology means better outcomes and the main task is managing risk. But we asked Aboriginal and Torres Strait Islander people how they are encountering AI in their everyday lives, and a different picture started to emerge. Our Relational Futures project explores Indigenous sovereignty and the governance of AI. Relational Futures positions AI not as a standalone tool, but as part of a wider system that shapes relationships between people, institutions, data and Country. We have now reported our findings, and there are clear warnings about what happens when questions of accountability, harm and care are ignored. As one participant told us, AI comes with “no accountability, no checks and balances, no responsibility”. Facing limited trust In Australia, we have seen automated decision-making lead to devastating consequences, such as in Robodebt. Similar dynamics are emerging in aged care and in the National Disability Insurance Scheme. These systems are often introduced in the name of efficiency. But efficiency for whom, and at what cost? AI and automated systems do not enter neutral environments. They enter institutions that already have uneven distributions of power, trust and accountability. When things go wrong, the impacts are not evenly felt. Our project set out to find the first qualitative baselines of Indigenous perspectives on AI, using surveys alongside yarning circles. We wanted to centre Indigenous perspectives and understand more deeply how Indigenous peoples experience new technologies. Our participants express limited trust in AI and, in many cases, a clear willingness to refuse using it. That refusal is not about rejecting technology outright. Participants recognised AI can intensify existing inequalities, particularly in sectors such as welfare, health and social services. There is a strong awareness that automation can make decisions faster – but also harder to see, harder to question, and harder to hold accountable. Understanding Indigenous data sovereignty Indigenous data sovereignty centres collective rights and responsibilities in the governance of data. It affirms the authority of Indigenous peoples to control data relating to their communities, lands and resources across the full data lifecycle. Such governance requires that data practices support self-determination, are grounded in community, and deliver collective benefit without reproducing harm or marginalisation. The participants in our research had a consistent emphasis on community benefit. The risks identified by our participants go well beyond privacy or data breaches. They pointed to environmental costs, the appropriation and flattening of Indigenous knowledges, and the lack of transparency in how systems are built and deployed. There is also a clear concern that AI will be used to fill gaps in under-resourced services. One participant said: There are times when AI doesn’t quite grasp the depth of First Nations experiences, cultural nuance or community dynamics. It can miss the emotional weight or the context, which reminds me that cultural authority must always sit with mob, not technology. An ‘AI Elder’ The project also pushed into more speculative territory, asking people to think about what AI could be, not just what it is now. One of the ideas we tested was an “AI Elder”, who could work in areas like reconnecting to culture, or providing advice on cultural matters. This was the ‘AI Elder’ we presented to our participants. Relational Futures We asked: what if AI was built around care, cultural knowledge, and responsibility to community, instead of speed and efficiency? But the reaction of our participants was blunt. Who would that Elder speak for? Who would it answer to? How could it have any real relationship to community? Elders aren’t just people who hold knowledge. They are part of community: they are trusted because of their relationships, their responsibilities, and their accountability over time. AI can’t be in relationship in that way, can’t be held accountable, can’t carry obligation. It can’t stand in connection to Country or community. Even when we try to imagine better versions of AI, there are some things that just don’t translate. A way forward AI governance cannot be limited to technical standards or compliance frameworks. It has to engage with authority, responsibility, harm and care. If AI systems can be designed in ways that are safe, accountable and beneficial for Aboriginal and Torres Strait Islander peoples – who are often the most surveilled and marginalised within systems – they are far more likely to be safe and effective for everyone. Designing for those at the margins is not a niche concern. It is a test of whether these systems work at all. As one participant told us: My biggest concern is that we get left behind. It’s easy to frame AI negatively, seeing it as a threat. It is just as easy to see the benefits it stands to offer. Clearly we need to be involved positively (we risk being left out otherwise) on how AI systems are designed, trained and used, otherwise there is a risk that existing power imbalances will be reproduced through technology. Relational Futures offers both a warning and a way forward. Without Indigenous leadership and relational approaches to governance, AI will continue to reproduce the kinds of harms already seen in systems like Robodebt. The way forward is less about slowing technology down, and more about rethinking what it is for, who it serves and how it is held to account. The authors do not work for, consult, own shares in or receive funding from any company or organisation that would benefit from this article, and have disclosed no relevant affiliations beyond their academic appointment.

These former MBB consultants left to found AI startups. Julius Bruch, Kevin Wu, Daphne Tay/ Bluente, Nathan Wangliao The path to partnership used to be the dream — especially at Bain, McKinsey, or BCG. But in the AI boom, some consultants have traded slide decks for startups. As AI reshapes industries and lowers the barriers to building a company, advising from the sidelines no longer feels like enough. Several of the former consultants we spoke to said they'd long had the itch to build. AI simply made the timing feel right. Others said that after years in consulting, owning a business — rather than recommending strategies to others — felt like the natural next step. The shift away from the Big Four — EY, Deloitte, PwC, and KPMG — has become more pronounced. James Ransome, a partner and strategy consulting lead at Patrick Morgan, a firm specializing in senior partner hiring and industry analysis, said last year that there was an "exodus" from traditional consulting powerhouses. At the same time, traditional consulting hiring has slowed. Data from workforce intelligence firm Revelio Labs shows entry-level inflows into consulting firms fell 54% year-over-year in June 2025. Among more senior hires — what Revelio classifies as "manager" level — monthly inflows dropped 22%, the firm told Business Insider in a September report. For those who moved, it wasn't easy to shed their consulting background. "I went from being in a huge office with all kinds of perks to being in a coworking space with just two other people," Nathan Wangliao, who left McKinsey in his 20s for startup life, told Business Insider. "We had to figure everything out from scratch." In his first month, he even considered going back. "This feels too scary," he recalled thinking. Many of these former consultants said they had to unlearn consulting habits, such as overanalyzing, chasing perfect data, and avoiding risk. As one of them put it, being in a startup means needing to "move fast with scrappy pragmatism." We spoke with four former consultants about why they left MBB firms to build AI startups, what consulting taught them, and what it took to break away. Read more stories about these consultants who left to become AI founders Read the original article on Business Insider

arXiv:2604.18725v1 Announce Type: new Abstract: The correlation between insect morphological traits and climate has been documented in physiological studies, but such studies remain limited by the time-consuming nature of the data analysis. In particular, the open source datasets often lack annotations of species' morphological traits, making dedicated annotations campaigns necessary; these efforts are typically local in scale and costly. In this paper, we propose a pipeline to identify and segment body parts of Odonates (dragonflies and damselflies) using deep neural networks, with the ultimate goal of extracting body parts' colouration. The pipeline is trained on a limited annotated dataset and refined with pseudo supervised data. We show that, by using open source images from citizen science platforms, our approach can segment each visible subject (Odonates) into head, thorax, abdomen, and wings and then extract a colour palette for each body part. This will enable large-scale statistical analysis of ecological correlations (e.g., between colouration and climate change, habitat loss, or geolocation) which are crucial for quantifying and assessing ecosystem biodiversity status.

arXiv:2604.18627v1 Announce Type: new Abstract: Ensuring human safety is of paramount importance in warehouse environments that feature mixed traffic of human workers and autonomous mobile robots (AMRs). Current approaches often treat humans as generic dynamic obstacles, leading to conservative AMR behaviors like slowing down or detouring, even when workers are fully aware and capable of safely sharing space. This paper presents a real-time vision-based method to estimate human awareness of an AMR using a single RGB camera. We integrate state-of-the-art 3D human pose lifting with head orientation estimation to ascertain a human's position relative to the AMR and their viewing cone, thereby determining if the human is aware of the AMR. The entire pipeline is validated using synthetically generated data within NVIDIA Isaac Sim, a robust physics-accurate robotics simulation environment. Experimental results confirm that our system reliably detects human positions and their attention in real time, enabling AMRs to safely adapt their motion based on human awareness. This enhancement is crucial for improving both safety and operational efficiency in industrial and factory automation settings.

arXiv:2604.18632v1 Announce Type: new Abstract: Stomata play a crucial role in regulating plant physiological processes and reflecting environmental responses. However, accurate and high-throughput stomatal phenotyping remains challenging, as conventional approaches rely on destructive sampling and manual annotation, restricting large-scale and field deployment. To overcome these limitations, a noninvasive restoration-detection integrated framework, termed StomaD2, is developed to achieve accurate and fast stomatal phenotyping under complex imaging conditions. The framework incorporates a diffusion-based restoration module to recover degraded images and a specialized rotated object detection network tailored to the small, dense, and cluttered characteristics of stomata. The proposed network enhances feature representation through three key innovations: a column-wise structure for global feature interaction, context-aware resampling and reweighting mechanism to improve multi-scale consistency, and a feature reassembly module to boost discrimination against complex backgrounds. In extensive comparisons, StomaD2 demonstrated state-of-the-art performance. On public Maize and Wheat datasets, it achieved accuracies of 0.994 and 0.992, respectively, significantly outperforming existing benchmarks. When benchmarked against ten other advanced models, including Oriented Former and YOLOv12, StomaD2 achieved a top-tier F1-score/mAP of 0.989. The framework is integrated into a user-friendly, field-operable system that supports the fast extraction of eight stomatal phenotypes, such as density and conductance. Validated on more than 130 plant species, StomaD2's results highlight its strong generalizability and potential for large-scale phenotyping, plant physiology analysis, and precision agriculture applications.

arXiv:2604.18648v1 Announce Type: new Abstract: Text-driven controllable dance generation remains under-explored, primarily due to the severe scarcity of high-quality datasets and the inherent difficulty of articulating complex choreographies. Characterizing dance is particularly challenging owing to its intricate spatial dynamics, strong directionality, and the highly decoupled movements of distinct body parts. To overcome these bottlenecks, we bridge principles from dance studies, human anatomy, and biomechanics to propose \textit{Choreographic Syntax}, a novel theoretical framework with a tailored annotation system. Grounded in this syntax, we combine professional dance archives with high-fidelity motion capture data to construct \textbf{DanceFlow}, the most fine-grained dance dataset to date. It encompasses 41 hours of high-quality motions paired with 6.34 million words of detailed descriptions. At the model level, we introduce \textbf{DanceCrafter}, a tailored motion transformer built upon the Momentum Human Rig. To circumvent optimization instabilities, we construct a continuous manifold motion representation paired with a hybrid normalization strategy. Furthermore, we design an anatomy-aware loss to explicitly regulate the decoupled nature of body parts. Together, these adaptations empower DanceCrafter to achieve the high-fidelity and stable generation of complex dance sequences. Extensive evaluations and user studies demonstrate our state-of-the-art performance in motion quality, fine-grained controllability, and generation naturalness.

arXiv:2604.18722v1 Announce Type: new Abstract: Cross-lingual transfer in NLP is often hindered by the ``script barrier'' where differences in writing systems inhibit transfer learning between languages. Transliteration, the process of converting the script, has emerged as a powerful technique to bridge this gap by increasing lexical overlap. This paper provides a comprehensive survey of the application of transliteration in cross-lingual NLP. We present a taxonomy of key motivations to utilize transliterations in language models, and provide an overview of different approaches of incorporating transliterations as input. We analyze the evolution and effectiveness of these methods, discussing the critical trade-offs involved, and contextualize their need in modern LLMs. The review explores various settings that show how transliteration is beneficial, including handling code-mixed text, leveraging language family relatedness, and pragmatic gains in inference efficiency. Based on this analysis, we provide concrete recommendations for researchers on selecting and implementing the most appropriate transliteration strategy based on their specific language, task, and resource constraints.

arXiv:2604.18758v1 Announce Type: new Abstract: Low-resource machine translation requires methods that differ from those used for high-resource languages. This paper proposes a novel in-context learning approach to support low-resource machine translation of the Coptic language to English, with syntactic augmentation from Universal Dependencies parses of input sentences. Building on existing work using bilingual dictionaries to support inference for vocabulary items, we add several representations of syntactic analyses to our inputs , specifically exploring the inclusion of raw parser outputs, verbalizations of parses in plain English, and targeted instructions of difficult constructions identified in sub-trees and how they can be translated. Our results show that while syntactic information alone is not as useful as dictionary-based glosses, combining retrieved dictionary items with syntactic information achieves significant gains across model sizes, achieving new state-of-the-art translation results for Coptic.

arXiv:2604.18740v1 Announce Type: new Abstract: Purpose: Automated C-arm positioning ensures timely treatment in patients requiring emergent interventions. When a conventional Deep Learning (DL) approach for C-arm control fails, clinicians must revert to manual operation, resulting in additional delays. Consequently, an agentic C-arm control framework based on multimodal large language models (MLLMs) is highly desirable, as it can incorporate clinician feedback and use reasoning to make adjustments toward more accurate positioning. Skeletal landmark localization is essential for C-arm control, and we investigate adapting MLLMs for autonomous landmark localization. Methods: We used an annotated synthetic X-ray dataset and a real X-ray dataset. Each X-ray in both datasets is paired with several skeletal landmarks. We fine-tuned two MLLMs and tasked them with retrieving the closest landmarks from each X-ray. Quantitative evaluations of landmark localization were performed and compared against a leading DL approach. We further conducted qualitative experiments demonstrating: (1) how an MLLM can correct an initially incorrect prediction through reasoning, and (2) how the MLLM can sequentially navigate the C-arm toward a target location. Results: On both datasets, fine-tuned MLLMs demonstrate competitive performance across all localization tasks when compared with the DL approach. In the qualitative experiments, the MLLMs provide evidence of reasoning and spatial awareness. Conclusion: This study shows that fine-tuned MLLMs achieve accurate skeletal landmark localization and hold promise for agentic autonomous C-arm control. Our code is available athttps://github.com/marszzibros/C-arm-localization-LLMs.git

arXiv:2604.18644v1 Announce Type: new Abstract: Predictive policing systems that allocate patrol resources based solely on predicted crime risk can unintentionally amplify racial disparities through feedback driven data bias. We present FASE, a Fairness Aware Spatiotemporal Event Graph framework, which integrates spatiotemporal crime prediction with fairness constrained patrol allocation and a closed loop deployment feedback simulator. We model Baltimore as a graph of 25 ZIP Code Tabulation Areas and use 139,982 Part 1 crime incidents from 2017 to 2019 at hourly resolution, producing a sparse feature tensor. The prediction module combines a spatiotemporal graph neural network with a multivariate Hawkes process to capture spatial dependencies and self exciting temporal dynamics. Outputs are modeled using a Zero Inflated Negative Binomial distribution, suitable for overdispersed and zero heavy crime counts. The model achieves a validation loss of 0.4800 and a test loss of 0.4857. Patrol allocation is formulated as a fairness constrained linear optimization problem that maximizes risk weighted coverage while enforcing a Demographic Impact Ratio constraint with deviation bounded by 0.05. Across six simulated deployment cycles, fairness remains within 0.9928 to 1.0262, and coverage ranges from 0.876 to 0.936. However, a persistent detection rate gap of approximately 3.5 percentage points remains between minority and non minority areas. This result shows that allocation level fairness constraints alone do not eliminate feedback induced bias in retraining data, highlighting the need for fairness interventions across the full pipeline.

arXiv:2604.18738v1 Announce Type: new Abstract: Masked diffusion language models such as LLaDA2.1 rely on Token-to-Token (T2T) editing to correct their own generation errors: whenever a different token crosses a confidence threshold, the committed token is overwritten. We identify three structural failure modes of this rule. The trigger cannot fire when no single alternative is confident enough; the replacement is computed under a context that may itself contain errors; and the uniform perturbations used to train the T2T stream do not resemble the coherent, semantically plausible mistakes that the model actually makes at inference. As an alternative, we propose Token-to-Mask (T2M) remasking. Rather than overwriting a suspect token with a new guess, T2M resets the position to the mask state, so that the next denoising step re-predicts it from an in-distribution context. The method is training-free, modifies only the editing rule, and introduces no new parameters. We pair it with three detection heuristics and give a short theoretical account of why a mask is a better conditioning signal than an erroneous token. Across 8 benchmarks, T2M improves accuracy on tasks that require exact token-level output. Its largest gain is +5.92 points on CMATH, where we attribute 79.9% of baseline errors to last-mile corruption (correct reasoning followed by a garbled final answer); T2M repairs 41.3% of these cases.

arXiv:2604.18744v1 Announce Type: new Abstract: Event cameras have recently shown promising capabilities in instantaneous motion estimation due to their robustness to low light and fast motions. However, computing wide-baseline correspondence between two arbitrary views remains a significant challenge, since event appearance changes substantially with motion, and learning-based approaches are constrained by both scalability and limited wide-baseline supervision. We therefore introduce the first event matching model that achieves cross-dataset wide-baseline correspondence in a zero-shot manner: a single model trained once is deployed on unseen datasets without any target-domain fine-tuning or adaptation. To enable this capability, we introduce a motion-robust and computationally efficient attention backbone that learns multi-timescale features from event streams, augmented with sparsity-aware event token selection, making large-scale training on diverse wide-baseline supervision computationally feasible. To provide the supervision needed for wide-baseline generalization, we develop a robust event motion synthesis framework to generate large-scale event-matching datasets with augmented viewpoints, modalities, and motions. Extensive experiments across multiple benchmarks show that our framework achieves a 37.7% improvement over the previous best event feature matching methods. Code and data are available at: https://github.com/spikelab-jhu/Match-Any-Events.

arXiv:2604.18713v1 Announce Type: new Abstract: Automated 3D segmentation of prostate lesions from biparametric MRI (bp-MRI) is essential for reliable algorithmic analysis, but achieving high precision remains challenging. Volumetric methods must combine multiple modalities while ensuring anatomical consistency, but current models struggle to integrate cross-modal information reliably. While vision-language models (VLMs) are replacing the currently used architectural designs, they still lack the fine-grained, lesion-level semantics required for effective localized guidance. To address these limitations, we propose a new multi-encoder U-Net architecture incorporating three key innovations: (1) an alignment loss that enhances foreground text-image similarity to inject lesion semantics; (2) a heatmap loss that calibrates the similarity map and suppresses spurious background activations; and (3) a final-stage, confidence-gated multi-head cross-attention refiner that performs localized boundary edits in high-confidence regions. A phase-scheduled training regime stabilizes the optimization of these components. Our method consistently outperforms prior approaches, establishing a new state-of-the-art on the PI-CAI dataset through enhanced multi-modal fusion and localized text guidance. Our code is available at https://github.com/NUBagciLab/Prostate-Lesion-Segmentation.

arXiv:2604.18753v1 Announce Type: new Abstract: An active challenge in developing multimodal machine learning (ML) models for healthcare is handling missing modalities during training and deployment. As clinical datasets are inherently temporal and sparse in terms of modality presence, capturing the underlying predictive signal via diagnostic multimodal ML models while retaining model explainability remains an ongoing challenge. In this work, we address this by re-framing clinical diagnosis as an autoregressive sequence modeling task, utilizing causal decoders from large language models (LLMs) to model a patient's multimodal trajectory. We first introduce a missingness-aware contrastive pre-training objective that integrates multiple modalities in datasets with missingness in a shared latent space. We then show that autoregressive sequence modeling with transformer-based architectures outperforms baselines on the MIMIC-IV and eICU fine-tuning benchmarks. Finally, we use interpretability techniques to move beyond performance boosts and find that across various patient stays, removing modalities leads to divergent behavior that our contrastive pre-training mitigates. By abstracting clinical diagnosis as sequence modeling and interpreting patient stay trajectories, we develop a framework to profile and handle missing modalities while addressing the canonical desideratum of safe, transparent clinical AI.

arXiv:2604.18751v1 Announce Type: new Abstract: Nonlinear machine-learning models are increasingly used to discover causal relationships in time-series data, yet the interpretation of their outputs remains poorly understood. In particular, causal scores produced by regularized neural autoregressive models are often treated as analogues of regression coefficients, leading to misleading claims of statistical significance. In this paper, we argue that causal relevance in nonlinear time-series models should be evaluated through forecast necessity rather than coefficient magnitude, and we present a practical evaluation procedure for doing so. We present an interpretable evaluation framework based on systematic edge ablation and forecast comparison, which tests whether a candidate causal relationship is required for accurate prediction. Using Neural Additive Vector Autoregression as a case study model, we apply this framework to a real-world case study of democratic development, modeled as a multivariate time series of panel data - democracy indicators across 139 countries. We show that relationships with similar causal scores can differ dramatically in their predictive necessity due to redundancy, temporal persistence, and regime-specific effects. Our results demonstrate how forecast-necessity testing supports more reliable causal reasoning in applied AI systems and provides practical guidance for interpreting nonlinear time-series models in high-stakes domains.

arXiv:2604.18729v1 Announce Type: new Abstract: Humor holds up a mirror to social perception: what we find funny often reflects who we are and how we judge others. When language models engage with humor, their reactions expose the social assumptions they have internalized from training data. In this paper, we investigate counterfactual unfairness through humor by observing how the model's responses change when we swap who speaks and who is addressed while holding other factors constant. Our framework spans three tasks: humor generation refusal, speaker intention inference, and relational/societal impact prediction, covering both identity-agnostic humor and identity-specific disparagement humor. We introduce interpretable bias metrics that capture asymmetric patterns under identity swaps. Experiments across state-of-the-art models reveal consistent relational disparities: jokes told by privileged speakers are refused up to 67.5% more often, judged as malicious 64.7% more frequently, and rated up to 1.5 points higher in social harm on a 5-point scale. These patterns highlight how sensitivity and stereotyping coexist in generative models, complicating efforts toward fairness and cultural alignment.

arXiv:2604.18592v1 Announce Type: new Abstract: We introduce a two-dimensional (2D) early exit strategy that coordinates layer-wise and sentence-wise exiting for classification tasks in large language models. By processing input incrementally sentence-by-sentence while progressively activating deeper layers, our method achieves multiplicative computational savings that exceed those from optimizing either dimension independently. Experimental evaluation across four state-of-the-art LLMs (Llama 3.1, Llama 3.2, Gemma, Qwen; 3B-8B parameters) on three sentiment classification datasets demonstrates additional speed-ups of 1.4--2.3$\times$ over optimal layer-wise early exit for simpler tasks with vanilla models, with graceful degradation on complex multi-class problems. Fine-tuning reduces but does not eliminate this advantage. The approach is model-agnostic, requires only lightweight classification adapters, and is orthogonal to complementary efficiency methods such as quantization and pruning. Our findings indicate that 2D early exit strategies excel when semantic information accumulates predictably across input structure, suggesting possible applicability to sequence-processing tasks beyond sentiment classification.

arXiv:2604.18739v1 Announce Type: new Abstract: Masked diffusion large language models (dLLMs) are a promising alternative to autoregressive generation. While reinforcement learning (RL) methods have recently been adapted to dLLM fine-tuning, their objectives typically depend on sequence-level marginal likelihoods, which are intractable for masked diffusion models. To address this, we derive Discrete Tilt Matching (DTM), a likelihood-free method that recasts dLLM fine-tuning as state-level matching of local unmasking posteriors under reward tilting. DTM takes the form of a weighted cross-entropy objective with explicit minimizer, and admits control variates that improve training stability. On a synthetic maze-planning task, we analyze how DTM's annealing schedule and control variates affect training stability and prevent mode collapse. At scale, fine-tuning LLaDA-8B-Instruct with DTM yields strong gains on Sudoku and Countdown while remaining competitive on MATH500 and GSM8K.